diff --git a/crates/ml-alpha/examples/alpha_rl_train.rs b/crates/ml-alpha/examples/alpha_rl_train.rs
index b625d5ef7..1b3b79251 100644
--- a/crates/ml-alpha/examples/alpha_rl_train.rs
+++ b/crates/ml-alpha/examples/alpha_rl_train.rs
@@ -255,6 +255,14 @@ struct Cli {
     /// diagnostic construction on most steps and let the GPU run free.
     #[arg(long, default_value = "1")]
     diag_every: usize,
+
+    /// Resume from checkpoint file path.
+    #[arg(long)]
+    resume_from: Option<PathBuf>,
+
+    /// Checkpoint interval in steps (default: 5000). Set to 0 to disable.
+    #[arg(long, default_value = "5000")]
+    checkpoint_every: usize,
 }
 
 fn parse_instrument_mode(s: &str) -> Result<InstrumentFilter, String> {
@@ -987,8 +995,17 @@ fn main() -> Result<()> {
     // and ~140ms of host-side Rust overhead per step.
     trainer.enable_mega_graph();
 
+    let start_step = if let Some(ref ckpt_path) = cli.resume_from {
+        let resumed_step = trainer.load_checkpoint(&dev, ckpt_path)
+            .with_context(|| format!("resume from {}", ckpt_path.display()))?;
+        eprintln!("resumed from checkpoint at step {resumed_step}");
+        resumed_step as usize
+    } else {
+        0
+    };
+
     let t_start = std::time::Instant::now();
-    for step in 0..cli.n_steps {
+    for step in start_step..cli.n_steps {
         // GPU-resident data loading: sample_and_gather + gather_next +
         // gather_current + gather_frd_labels all run as GPU kernels
         // inside step_with_lobsim_gpu. Zero CPU data work per step.
@@ -1111,6 +1128,18 @@ fn main() -> Result<()> {
         // blocks on diag I/O.
         let _ = diag_tx.try_send(frame);
 
+        if cli.checkpoint_every > 0 && step > 0 && step % cli.checkpoint_every == 0 {
+            let ckpt_path = cli.out.join(format!("checkpoint-{step}.bin"));
+            trainer.save_checkpoint(&ckpt_path, step as u64)
+                .with_context(|| format!("checkpoint at step {step}"))?;
+            eprintln!("checkpoint saved: {}", ckpt_path.display());
+            let old_step = step.saturating_sub(cli.checkpoint_every * 2);
+            if old_step > 0 {
+                let old_path = cli.out.join(format!("checkpoint-{old_step}.bin"));
+                let _ = std::fs::remove_file(&old_path);
+            }
+        }
+
         if step % cli.log_every == 0 || step == cli.n_steps - 1 {
             let isv = trainer.isv_host_slice();
             let elapsed = t_start.elapsed().as_secs_f32();